| ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py | |
| ; RUN: llc -mtriple=amdgpu9.00-amd-amdhsa < %s | FileCheck --check-prefix=GCN %s | |
| define weak_odr amdgpu_kernel void @test_mul24_knownbits_kernel(ptr addrspace(1) %p) #4 { | |
| ; GCN-LABEL: test_mul24_knownbits_kernel: | |
| ; GCN: ; %bb.0: ; %entry | |
| ; GCN-NEXT: v_and_b32_e32 v0, 3, v0 | |
| ; GCN-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x0 | |
| ; GCN-NEXT: v_mul_i32_i24_e32 v0, -5, v0 | |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffffffe0, v0 | |
| ; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NEXT: v_lshlrev_b64 v[0:1], 2, v[0:1] | |
| ; GCN-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NEXT: v_mov_b32_e32 v2, s1 | |
| ; GCN-NEXT: v_add_co_u32_e32 v0, vcc, s0, v0 | |
| ; GCN-NEXT: v_addc_co_u32_e32 v1, vcc, v2, v1, vcc | |
| ; GCN-NEXT: v_mov_b32_e32 v2, 0 | |
| ; GCN-NEXT: global_store_dword v[0:1], v2, off | |
| ; GCN-NEXT: s_endpgm | |
| entry: | |
| %0 = tail call i32 @llvm.amdgcn.workitem.id.x() #28, !range !4 | |
| %tid = and i32 %0, 3 | |
| %1 = mul nsw i32 %tid, -5 | |
| %v1 = and i32 %1, -32 | |
| %v2 = sext i32 %v1 to i64 | |
| %v3 = getelementptr inbounds float, ptr addrspace(1) %p, i64 %v2 | |
| store float 0.000, ptr addrspace(1) %v3, align 4 | |
| ret void | |
| } | |
| define i32 @f(i32 %x, i32 %y) { | |
| ; GCN-LABEL: f: | |
| ; GCN: ; %bb.0: | |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) | |
| ; GCN-NEXT: v_or_b32_e32 v0, 0xffff80, v0 | |
| ; GCN-NEXT: v_or_b32_e32 v1, 0xffff80, v1 | |
| ; GCN-NEXT: v_mul_i32_i24_e32 v0, v0, v1 | |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 14, v0 | |
| ; GCN-NEXT: s_setpc_b64 s[30:31] | |
| %xx = or i32 %x, -128 ; 0xffffff80 | |
| %yy = or i32 %y, -128 ; 0xffffff80 | |
| %r = mul i32 %xx, %yy | |
| %rr = lshr i32 %r, 14 | |
| ret i32 %rr | |
| } | |
| ; The LHS is completely unknown, but the constant RHS still provides known | |
| ; bits for the product: 12 == 4 * 3, so the low 2 bits of the result are 0. | |
| define i32 @mul_i24_known_rhs_unknown_lhs(i32 %x) { | |
| ; GCN-LABEL: mul_i24_known_rhs_unknown_lhs: | |
| ; GCN: ; %bb.0: | |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) | |
| ; GCN-NEXT: v_mov_b32_e32 v0, 0 | |
| ; GCN-NEXT: s_setpc_b64 s[30:31] | |
| %mul = call i32 @llvm.amdgcn.mul.i24.i32(i32 %x, i32 12) | |
| %and = and i32 %mul, 3 | |
| ret i32 %and | |
| } | |
| define i32 @mul_u24_known_rhs_unknown_lhs(i32 %x) { | |
| ; GCN-LABEL: mul_u24_known_rhs_unknown_lhs: | |
| ; GCN: ; %bb.0: | |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) | |
| ; GCN-NEXT: v_mov_b32_e32 v0, 0 | |
| ; GCN-NEXT: s_setpc_b64 s[30:31] | |
| %mul = call i32 @llvm.amdgcn.mul.u24.i32(i32 %x, i32 12) | |
| %and = and i32 %mul, 3 | |
| ret i32 %and | |
| } | |
| ; The known value is the LHS and the RHS is unknown. The known bits of the LHS | |
| ; must still be used. | |
| define i32 @mul_i24_known_lhs_unknown_rhs(i32 %x) { | |
| ; GCN-LABEL: mul_i24_known_lhs_unknown_rhs: | |
| ; GCN: ; %bb.0: | |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) | |
| ; GCN-NEXT: v_mov_b32_e32 v0, 0 | |
| ; GCN-NEXT: s_setpc_b64 s[30:31] | |
| %mul = call i32 @llvm.amdgcn.mul.i24.i32(i32 12, i32 %x) | |
| %and = and i32 %mul, 3 | |
| ret i32 %and | |
| } | |
| ; The known value is the LHS and the RHS is unknown, for the unsigned | |
| ; multiply. | |
| define i32 @mul_u24_known_lhs_unknown_rhs(i32 %x) { | |
| ; GCN-LABEL: mul_u24_known_lhs_unknown_rhs: | |
| ; GCN: ; %bb.0: | |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) | |
| ; GCN-NEXT: v_mov_b32_e32 v0, 0 | |
| ; GCN-NEXT: s_setpc_b64 s[30:31] | |
| %mul = call i32 @llvm.amdgcn.mul.u24.i32(i32 12, i32 %x) | |
| %and = and i32 %mul, 3 | |
| ret i32 %and | |
| } | |
| ; A self-multiply of a defined but otherwise unknown value still has bit 1 | |
| ; clear: x * x is 0 or 1 modulo 4. Both operands are the same value, so the | |
| ; known bits of the shared operand are computed once and used for both sides. | |
| define i32 @mul_i24_self_square_freeze(i32 %x) { | |
| ; GCN-LABEL: mul_i24_self_square_freeze: | |
| ; GCN: ; %bb.0: | |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) | |
| ; GCN-NEXT: v_mov_b32_e32 v0, 0 | |
| ; GCN-NEXT: s_setpc_b64 s[30:31] | |
| %f = freeze i32 %x | |
| %mul = call i32 @llvm.amdgcn.mul.i24.i32(i32 %f, i32 %f) | |
| %and = and i32 %mul, 2 | |
| ret i32 %and | |
| } | |
| define i32 @mul_u24_self_square_freeze(i32 %x) { | |
| ; GCN-LABEL: mul_u24_self_square_freeze: | |
| ; GCN: ; %bb.0: | |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) | |
| ; GCN-NEXT: v_mov_b32_e32 v0, 0 | |
| ; GCN-NEXT: s_setpc_b64 s[30:31] | |
| %f = freeze i32 %x | |
| %mul = call i32 @llvm.amdgcn.mul.u24.i32(i32 %f, i32 %f) | |
| %and = and i32 %mul, 2 | |
| ret i32 %and | |
| } | |
| ; A self-multiply of a known odd value: x is odd so bit 0 is 1 and the square | |
| ; of an odd number has bits 1 and 2 clear, i.e. the low bits are 001. | |
| ; | |
| ; The mask is narrowed from 7 to 1 here, which shows that bits 1 and 2 are | |
| ; known zero. SelectionDAG does not fold the remaining "and 1" to the constant | |
| ; 1 because AMDGPUTargetLowering::SimplifyDemandedBitsForTargetNode does not | |
| ; report the known bits of target opcodes during the demanded-bits recursion | |
| ; (unlike the base implementation used by other targets), so the operand looks | |
| ; unknown to the fold. The GlobalISel value tracking test checks the full | |
| ; 001 result. | |
| define i32 @mul_i24_self_square_odd(i32 %x) { | |
| ; GCN-LABEL: mul_i24_self_square_odd: | |
| ; GCN: ; %bb.0: | |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) | |
| ; GCN-NEXT: v_or_b32_e32 v0, 1, v0 | |
| ; GCN-NEXT: v_mul_i32_i24_e32 v0, v0, v0 | |
| ; GCN-NEXT: v_and_b32_e32 v0, 1, v0 | |
| ; GCN-NEXT: s_setpc_b64 s[30:31] | |
| %f = freeze i32 %x | |
| %o = or i32 %f, 1 | |
| %mul = call i32 @llvm.amdgcn.mul.i24.i32(i32 %o, i32 %o) | |
| %and = and i32 %mul, 7 | |
| ret i32 %and | |
| } | |
| ; A 24x24 product needs at most 48 bits, so the upper 16 bits of the i64 | |
| ; result are known zero even when both operands are unknown. | |
| define i64 @mul_u24_i64_upper_bits_zero(i32 %x, i32 %y) { | |
| ; GCN-LABEL: mul_u24_i64_upper_bits_zero: | |
| ; GCN: ; %bb.0: | |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) | |
| ; GCN-NEXT: v_mov_b32_e32 v0, 0 | |
| ; GCN-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NEXT: s_setpc_b64 s[30:31] | |
| %mul = call i64 @llvm.amdgcn.mul.u24.i64(i32 %x, i32 %y) | |
| %hi = lshr i64 %mul, 48 | |
| ret i64 %hi | |
| } | |
| ; Function Attrs: nounwind readnone speculatable | |
| declare i32 @llvm.amdgcn.workitem.id.x() #20 | |
| declare i32 @llvm.amdgcn.mul.i24.i32(i32, i32) | |
| declare i32 @llvm.amdgcn.mul.u24.i32(i32, i32) | |
| declare i64 @llvm.amdgcn.mul.u24.i64(i32, i32) | |
| !4 = !{i32 0, i32 1024} |